Context Combined Cascaded Region Proposal Network with Dual Loss Function for Accurate Car Detection

Car detection in the wild suffers from the problem of high occlusion and complicated scene disturbance. In this paper, we propose a novel Context Combined Cascaded Region Proposal Network (C3RPN) with dual loss function which addresses these problems at three dimensions: (i) we propose a cascaded two-stage region proposal structure that removes the limitation of fixed anchor boxes, (ii) Smooth L1 loss and Intersection over Union (IoU) loss are adopted to optimize the two stages respectively for better convergence and finer proposal regression, (iii) we integrate the complementary information between the proposed area and its context to strengthen the ability of accurate localization. We conduct extensive experiments on KITTI benchmark to prove the effectiveness of C3RPN. Results show that the proposed C3RPN outperforms Faster-RCNN with a large margin of 10.91% relative mAP gain and is robust to occlusion and background variation.